distributed-tracing

Deploy the Jaeger Operator into an observability namespace, then apply a Jaeger custom resource with a production strategy and Elasticsearch storage. For local development, run the jaegertracing/all-in-one Docker image exposing the UI on port 16686.

Create a TracerProvider with your service name, add a BatchSpanProcessor with a JaegerExporter pointing at the agent host, and call FlaskInstrumentor().instrument_app(app). Then create custom spans with tracer.start_as_current_span around business logic and database calls.

Jaeger provides a dedicated UI with service dependency graphs and supports Elasticsearch storage. Tempo integrates with Grafana and stores traces in object storage like S3, accepting Jaeger, Zipkin, and OTLP protocols. Both work with OpenTelemetry instrumentation.

Missing traces usually come from a wrong collector endpoint, network connectivity issues, or sampling configuration dropping all spans. Check the exporter endpoint, verify the agent ports are reachable, and review application logs for exporter errors.

Sample 1-10% of traces in production to keep overhead under 1% CPU impact. Use probabilistic sampling with a param like 0.01, rate-limiting sampling for a fixed traces-per-second cap, or parent-based sampling for deterministic decisions.